Applied AI · Retrieval · LLM Systems
I build AI systems end to end: retrieval pipelines, LLM tooling, and the infrastructure that makes them hold up in real use.
| Project | What it does |
|---|---|
| 📦 isotrieve |
Migrate a vector database to a new embedding model without re-embedding the corpus. Fits a linear transform from ~2K calibration texts and gates the migration on measured retrieval retention (87–91% on BEIR). Adapters for Chroma, Qdrant, LangChain, and LlamaIndex. |
| ⚡ FlashGPT | GPT-style language model built from scratch in PyTorch, with distillation, distributed training, and Tree-of-Thought reasoning modules. |
| 🤖 Persona Chatbot | Fine-tuned GPT-3.5 and Llama 3 for persona consistency on Google's Synthetic Persona Chat, evaluated with BERTScore, ROUGE, and perplexity. |
| 🦙 LlamaPy | Llama 3 implemented from scratch in Python and PyTorch. |
| 🔧 Finetune-Llama3 | Reusable templates for fine-tuning Llama 3 models with Unsloth. |
| 🌐 Web-Serve | High-performance, configurable HTTP server in C++ (Boost, CMake, Docker, GoogleTest). |
- 🏎️ Reinforcement Learning: DQN and DDQN agents for a car-racing environment in Gymnasium and PyTorch
- 🧠 SynaptoNet-CRNN: EEG signal classification with CNN, LSTM, and CNN+LSTM architectures
- 🎾 Tennis Match Viewer: match playback app that lets coaches filter and search points for film review (React, Firebase)
- 🃏 Texas Hold'em on FPGA: full poker game implemented in Verilog on a Basys3 board
- 💾 Ext2 File System: ext2 implementation with mounting, symlinks, bitmaps, and inodes in C
Languages: Python • TypeScript • JavaScript • C/C++ • Java • SQL • Verilog
AI / ML: PyTorch • Hugging Face Transformers • scikit-learn • Unsloth • RAG pipelines • Vector databases (Qdrant, Chroma, pgvector) • MCP
Web & Infra: React • Next.js • Node.js • Flask • Django • AWS (Lambda, API Gateway, SQS) • Docker • PostgreSQL • MongoDB • Firebase


